Training data for an activation oracle (AO): a reader that recovers a subject
model's committed-but-not-yet-stated intent directly from its activations. Each
row is a transcript cut mid-generation; the label is a latent quantity the
subject (Qwen/Qwen3-8B, NOTHINK chat markup) has already decided but has
not yet written into the visible text.
92,478 rows across two build batches (in1 7,096 · in2 85,382), spanning
7 components… See the full description on the dataset page:
https://huggingface.co/datasets/cds-jb/intent-qwen3-8b-train.
